Pointing Gesture Classification Dataset (MediaPipe Hand Landmark Distances)

This dataset is specifically designed to support machine learning/deep learning tasks for pointing gesture classification. It contains 13,575 annotated instances of hand gestures extracted from children’s pointing, acquired through RGB video frames. Each sample is represented by 20 numerical features corresponding to the Euclidean distances of MediaPipe hand landmarks (index 1 to 20) from the wrist (landmark 0), which serves as the reference point (value = 0 for all samples). This normalized representation allows for robust comparison across different hand sizes, orientations, and subjects.  The dataset is structured into two binary classes:

0 – No Pointing: frames where no pointing gesture is being performed.

1 – Pointing: frames where a pointing gesture is clearly identified.

Each row in the CSV file corresponds to a single video frame and includes:

20 continuous features: landmark distances from the wrist (points 1–20 as per MediaPipe’s anatomical indexing).

1 target label: Label (0 or 1), indicating the gesture class.

The landmark indices follow the official MediaPipe hand model definition:

0: WRIST

1-4: THUMB (CMC, MCP, IP, TIP)

5-8: INDEX_FINGER (MCP, PIP, DIP, TIP)

9-12: MIDDLE_FINGER (MCP, PIP, DIP, TIP)

13-16: RING_FINGER (MCP, PIP, DIP, TIP)

17-20: PINKY (MCP, PIP, DIP, TIP)